- 1Say what done meansIn the first sentence
- 2It verifies itselfReal commands, real UI, evidence
- 3Open a PRSmall, ordered commits
- 4Babysit to greenConflicts, review comments, CI
- 5MergeYou say land the stackNeeds you
Say what done means in the first sentence
/pstack:poteto-mode Add JSON output to this command. The text output must stay byte-for-byte identical, the JSON must parse, and run both on the example project. Show me the evidence.The reply should include the real commands and their output. If a check couldn’t run, it says “inconclusive”.
“It compiles” isn’t proof. Ask for evidence that fits the change:
| What changed | Evidence to ask for |
|---|---|
| Command line | Run the real command |
| UI | Walk through it in the running app |
| Parsers, migrations | Replay a saved input |
| Performance | Compare before and after measurements |
| Storage | Read back the value written |
Let the AI verify in your app
If the AI has no way to drive your app, run this first:
/pstack:create-verification-skillWhat you get: a /verify-<your-app-name> verification skill. It mostly reads your code, and does a full trial run before handing it over. If the trial run fails, don’t use what it made.
After that, ask for evidence like this:
/pstack:poteto-mode Build the bulk archive feature. Verify it with /verify-<your-app-name>, and give me a screen recording and screenshots as evidence.Commit the verification skill to the repo. Review it regularly with /pstack:maintain-verification-skill (daily is best). For what it generates, how to check it, and how the author uses it, see Give your project a verification skill.
Open the PR
A PR is a set of changes sent for review, then merged.
/pstack:poteto-mode Open a PR. Small, ordered commits, with the evidence in the description.What you get: it arranges small commits in a worktree, and gives you the PR link at the end.
Get it ready to merge: babysit the PR
/pstack:poteto-mode Babysit this PR and get it green.What you get: it handles conflicts, then review comments, then CI (the checks that run after every push). Real problems get fixed. Comments without merit get a reply with the reason.
It stops at “ready to merge”. It never merges for you.Merge: land the stack
A stack is a chain of PRs built on top of each other.
/pstack:poteto-mode land the stack.What you get: an AI that didn’t write each PR verifies it again. Starting from the bottom, it merges only the unbroken run that passed. It tells you where the chain broke.
Stuck?
It says it's done, but didn't show any evidence.
Reply: “apply prove it works. Show me the real output, not the build log.” prove it works means “prove it works with real results”. It’s a principle name pstack knows.
I measured a speedup. How do I know it's real?
/pstack:benchmark-checklist Check this export speedup before it goes into the PRIt asks 7 questions. There are only four verdicts: faster, slower, no measurable difference, inconclusive.
It's a small change, but I'm uneasy.
Use /pstack:blast-radius. It finds why the change is safe, then proves it by running code.
Some cases tests can't reach, like two retries colliding.
Use formal verification: /pstack:create-formal-verification-skill Task leases and retry logic. The same task must never run twice at the same time. When it’s worth it and how to read a counterexample: Bugs tests can’t reach: formal verification.
Before moving on, check
✓ All ticked — on to the next step.